Skip to content


Use Large Language models with workflows. With this block in place, one may prompt both GPT-4V and CogVLM models and combine their outputs with other workflow components, effectively building powerful applications without single line of code written.

The LMM block allows to specify structure of expected output, automatically inject the specification into prompt and parse expected structure into block outputs that are accessible (and can be referred) by other workflows components. LMMs may occasionally produce non-parsable results according to specified output structure - in that cases, outputs will be filled with not_detected value.

Step parameters

  • type: must be LMM (required)
  • name: must be unique within all steps - used as identifier (required)
  • image: must be a reference to input of type InferenceImage or crops output from steps executing cropping ( Crop, AbsoluteStaticCrop, RelativeStaticCrop) (required)
  • prompt: must be string of reference to InferenceParameter - value holds unconstrained text prompt to LMM model (required).
  • lmm_type: must be string of reference to InferenceParameter - value holds the type of LMM model to be used - allowed values: gpt_4v and cog_vlm (required)
  • lmm_config: (optional) structure that has the following schema:
      "max_tokens": 450,
      "gpt_image_detail": "low",
      "gpt_model_version": "gpt-4-vision-preview"
    to control inner details of LMM prompting. All parameters now are suited to control GPT API calls. Default for max tokens is 450, gpt_image_detail default is auto (allowed values: low, auto, high), gpt_model_version is gpt-4-vision-preview.
  • remote_api_key - optional string or reference to InferenceParameter that holds API key required to call LMM model - in current state of development, we require OpenAI key when lmm_type=gpt_4v and do not require additional API key for CogVLM calls.
  • json_output: optional dict[str, str] (pointing expected output JSON field name to its description) or reference to InferenceParameter with such dict. This field is used to instruct model on expected output format. One may not specify field names: ["raw_output", "structured_output", "image", "parent_id"], due to the fact that keys from json_output dict will be registered as block outputs (to be referred by other blocks) and cannot collide with basic outputs of that block. Additional outputs **will only be registered if defined in-place, not via InferenceParameter).

Step outputs

  • raw_output - raw output of LMM for each input image
  • structured_output - if json_output is specified, whole parsed dictionary for each input image will be placed in this field, otherwise for each image, empty dict will be returned
  • image - size of input image, that predictions coordinates refers to
  • parent_id - identifier of parent image / associated detection that helps to identify predictions with RoI in case of multi-step pipelines
  • for each of json_output - dedicated field will be created (with values provided image-major) - and those can be referred as normal outputs ($steps.{step_name}.{field_name}).

Important notes

  • CogVLM can only be used in self-hosted API - as Roboflow platform does not support such model. Use inference server start on a machine with GPU to test that model.